Accurate measure of the risk aversion of farmers.

Why?

• pesticides -> lower risk of loss from farming

• pesticides applied before pests arrive

• if farming is a lottery, pesticides reduce its volatility

• the higher the risk (loss?) aversion, the higher the pesticide use

with a precise risk aversion measure…

1. estimation of subjective costs to farmers of cutting or eliminating pesticides;
2. amount of pesticides that could be cut by imposing the risk-neutral pesticide use;
3. potential adjustment of policy to individual (or sector) risk attitudes.

State of the art in risk literature is far from optimal

• despite experimental and empirical efforts, there are problems with the very concept of risk attitudes
• experimental methods currently used have shown low external validity
• low correlation with questionnaires & field behavior

Developing and testing in the field a new risk elicitation measure

• takes into account noise & cognitive abilities

• takes into account risk perception

• theory and lab experiments in progress (ANR RETRISK)

• FAST: application to the field and farmers

A meta-analysis of Risk elicitation tasks

• elicited risk atitudes: tasks and questionnaires

• convergent validity: correlation among tasks

• convergent validity: correlation among questionnaires

• predictive validity: correlation task \(\iff\) questionnaires

• preregistration on OSF

• transparent data collection & analysis on gitHub

• live data exploration on a shiny app

• Gnambs Appel and Oeberst (PONE 2015)
• Crosetto and Filippin (EXEC 2016)
• Filippin and Crosetto (ManSci 2016)
• Pedroni Frey Bruhin Dutilh Hertwig and Rieskamp (NHB 2016)
• Menkhoff and Sakha (JEconPsy 2017)
• Frey Pedroni Mata Rieskamp and Hertwig (ScAdv 2017)
• Nielsen (JEBO 2019)
• Charness Garcia Offerman and Villeval (WP 2019)
• Holzmeister and Stefan (WP 2018)
• Zhou and Hey (ExEc 2018)
• Fairley Parelman Jones and McKell Carter (JEconPsy 2018)
• Csermely Rabas (JRU 2018)

\(u(x) = x^r\)

• simple
• captures risk aversion
• makes different tasks comparable

A difficult task with crucial relevance

• directly unobservable
• latent construct (\(\Rightarrow\) requires a theory)
• should we..
  • infer from real world data or from ad-hoc choices
  • ask or task?
  • elicit by descrption or by experience?

and by the way, what is risk?

The act of implementing a goal-directed option qualifies as an instance of risk taking whenever two things are true: (a) the behavior in question could lead to more than one outcome and (b) some of these outcomes are undesirable or even dangerous. In essence, then, risk taking involves the implementation of options that could lead to negative consequences.

(Byrnes et al 1999)

risk loosely defined as probability of harm

focus on questionnaires and intuitive tasks

• Quests:
  • directly ask
  • over different domains
  • tackle risk perception
• Tasks
  • putting the subject in a ‘risky’ situation
  • card/gambling tasks

Metrics of success: convergent validity + predictive validity

decisions given a probability distribution over outcomes

• if probability and outcomes known: risk

• if only oucomes known: ambiguity

• if both unknown: knightian uncertainty

risk formally defined as uncertainty over outcomes

focus on decontextualized tasks (and questionnaires)

• The lottery paradigm
  • incentives
  • risk task = choice over lotteries
  • different formats, cover stories, contexts
  • strong theoretical underpinning
  • estimation of utility functions (\(\Rightarrow\) models)

Metric of success: internal validity (task \(\iff\) theory)

• low correlations with questionnaires

• across questionnaires and tasks

• Beauchamp et al JRU 2016: questionnaires are rather predictive

• task-specific bias

• (noise)

• risk perception

• (theory)

noisy preference + one-shot choices \(\Rightarrow\) noisy data

• cognitive limits \(\Rightarrow\) limited understanding

• task-specific bias?

(this work: Crosetto and Filippin, ExEc 2015)

How does the mere mechanics of each task affect the outcome?

Simulation exercise:

• generate 100k virtual agents
• for each agent, \(r\sim N(0.7,0.3)\)
• let the agents play each of the 4 tasks
• collect results, run statistics
• analyze the retrieved \(\hat{r}\)

A good task should be able to recreate the starting distribution, if no error.

3 types of simulations:

• deterministic

• random parameter model \(\Rightarrow\) models fuzzy preferences

  • for each agent, \(r = r_0 + \varepsilon, \; \varepsilon \sim \mathcal{N}(0,\mu)\)
  • \(\mu \in (0.3;0.6)\)

• random agents \(\Rightarrow\) models frame effects

  • 10% of subjects act randomly on the space of the task

is there a task-specific bias? yes

does it account for all differences? no

is this the only way to take noise into account? no

• economists assume subjects share the same risk definition

• namely:

  • risk as a distribution of probability over outcomes
  • \(EV\) as the average across all possible states of the world
  • risk aversion as diminishing marginal utility of money
  • subjects care about variance

• but subjects think of risk as probability of a loss

• do subjects find our tasks risky?

• We do not know because we assume they do

• Holzmeister et al Working Paper
• gave description of return from an asset to subjects
• \(\sim\) 7000 subjects
• including \(\sim\) 2500 traders
• asked to rate perceived risk of each asset